Tiered Reasoning for Intuitive Physics: Toward Verifiable Commonsense Language Understanding

Abstract

Large-scale, pre-trained language models (LMs) have achieved human-level performance on a breadth of language understanding tasks. However, evaluations only based on end task performance shed little light on machines' true ability in language understanding and reasoning. In this paper, we highlight the importance of evaluating the underlying reasoning process in addition to end performance. Toward this goal, we introduce Tiered Reasoning for Intuitive Physics (TRIP), a novel commonsense reasoning dataset with dense annotations that enable multi-tiered evaluation of machines' reasoning process. Our empirical results show that while large LMs can achieve high end performance, they struggle to support their predictions with valid supporting evidence. The TRIP dataset and our baseline results will motivate verifiable evaluation of commonsense reasoning and facilitate future research toward developing better language understanding and reasoning models.

Benchmarks for Automated Commonsense Reasoning: A Survey

“…113,000 questions Crowd sourcing TG-CSR [116] Question answering 331 questions Expert construction TimeDial [109] Cloze tasks for 1,100 dialogues Crowd sourcing temporal reasoning Torque [103] Order of events 3,200 news stories Crowd sourcing in a news story 21,000 questions Tracie [161] Order of implicit 5500 problems Crowd sourcing and explicit event Triangle COPA [51] Why did that thing 100 examples Expert construction do that? TRIP [129] Which story is more 2100 stories Crowd sourcing plausible? WNLI [136] Entailment 849 sentences Expert construction.…”